See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/220605190

Computer Models of Creativity

Article  in  Ai Magazine · July 2009

DOI: 10.1609/aimag.v30i3.2254 · Source: DBLP

CITATIONS

178

1 author:

Margaret A. Boden
University of Sussex
115 PUBLICATIONS   3,155 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Margaret A. Boden on 25 September 2014.

The user has requested enhancement of the downloaded file.


Computer Models
of Creativity
n Creativity isn’t magical. It’s an aspect of nor-
mal human intelligence, not a special faculty
granted to a tiny elite. There are three forms:
combinational, exploratory, and transforma-
tional. All three can be modeled by AI—in some
cases, with impressive results. AI techniques
underlie various types of computer art. Whether
computers could “really” be creative isn’t a sci-
entific question but a philosophical one, to
which there’s no clear answer. But we do have
the beginnings of a scientific understanding of
creativity.
Copyright © 2009, Association for the Advancement of Artificial Intelligence. All rights reserved. ISSN 0738-4602
Articles
Margaret A. Boden
C
reativity is a marvel of the human mind, and an obvious
goal for AI workers. Indeed, the proposal that led to the famous
1956 Dartmouth Summer School often remembered as the time
of AI’s birth mentioned “creativity,” “invention,” and “discov-
ery” as key aims for the newly named discipline (McCarthy et
al. 1955, 45, 49ff.). And, 50 years later, Herb Simon—in an e-
mail discussion between AAAI Fellows (quoted in Boden 2006,
1101)—cited a paper on creativity (Simon 1995) in answer to
the challenge of whether AI is a science, as opposed to “mere”
engineering.
But if its status as an AI goal is obvious, its credibility as a
potential AI achievement is not. Many people, including many
otherwise hard-headed scientists, doubt—or even deny out-
right—the possibility of a computer’s ever being creative.
Sometimes, such people are saying that, irrespective of its per-
formance (which might even match superlative human exam-
ples), no computer could “really” be creative: the creativity lies
entirely in the programmer. That’s a philosophical question that
needn’t detain us here (but see the final section, below).
More to the point, these people typically claim that a com-
puter simply could not generate apparently creative perform-
ance. That’s a factual claim—which, in effect, dismisses AI
research on creativity as a waste of time.
However, it is mistaken: AI scientists working in this area
aren’t doomed to disappointment. It doesn’t follow that they
will ever, in practice, be able to engineer a new Shakespeare or
a neo-Mozart (although the latter goal has been approached
more nearly than most people imagine). But lesser examples of
AI creativity already abound. And, crucially, they help us to
understand how human creativity is possible.
FALL 2009 23
Articles
What Is Creativity?
Creativity can be defined as the ability to generate
novel, and valuable, ideas. Valuable, here, has
many meanings: interesting, useful, beautiful, sim-
ple, richly complex, and so on. Ideas covers many
meanings too: not only ideas as such (concepts,
theories, interpretations, stories), but also artifacts
such as graphic images, sculptures, houses, and jet
engines. Computer models have been designed to
generate ideas in all these areas and more (Boden
2004).
As for novel, that has two importantly different
meanings: psychological and historical. A psycho-
logical novelty, or P-creative idea, is one that’s new
to the person who generated it. It doesn’t matter how
many times, if any, other people have had that
idea before. A historical novelty, or H-creative idea,
is one that is P-creative and has never occurred in
history before.
So what we need to explain, here, is P-creativi-
ty—which includes H-creativity but also covers
more mundane examples. And our explanation
must fit with the fact that creativity isn’t a special
faculty, possessed only by a tiny Romantic elite.
Rather, it’s a feature of human intelligence in gen-
eral. Every time someone makes a witty remark,
sings a new song to his or her sleepy baby, or even
appreciates the topical political cartoons in the dai-
ly newspaper, that person is relying on processes of
creative thought that are available to all of us.
To be sure, some people seem to be better at it
than others. Some newspaper cartoonists have an
especially good eye, and brain, for the delectable
absurdities of our political masters. And a few peo-
ple come up with highly valued H-creative ideas
over and over again. Alan Turing is one example
(he did revolutionary work in mathematics, com-
puter science, cryptography, and theoretical biolo-
gy [Boden 2006, 3.v.b–d, 4.i–ii, 15.iv]). But some
people are better at tennis, too. To understand how
Wimbledon champions manage to do what they
do, one must first understand how Jo Bloggs can
do what he does at the municipal tennis courts. P-
creativity, whether historically novel or not, is
therefore what we must focus on.
Computer models sometimes aim for, and even
achieve, H-creativity. For example, a quarter cen-
tury ago, an AI program designed a three-dimen-
sional silicon chip that was awarded a patent—
which requires that the invention must not be
“obvious to a person skilled in the art” (Lenat
1983). And the AARON program (mentioned
below) that generates beautifully colored drawings
is described by its human originator as a “world-
class” colorist. So it’s presumably H-creative—and
it’s certainly capable of coming up with color
schemes that he himself admits he wouldn’t have
had the courage to use.
Often, however, computer models aim merely
24 AI MAGAZINE
for P-creativity. Examples discussed below include
drawing scientific generalizations that were first
discovered centuries ago (Langley et al. 1987), or
generating music of a type composed by long-dead
musicians (Cope 2001, 2006)
Even P-creativity in computers need not match
all the previous achievements of human beings.
Years ago, in the early days of AI, Seymour Papert
used to warn AI researchers, and their sceptical crit-
ics, against “the superhuman human fallacy.” That
is, we shouldn’t say that AI has failed simply
because it can’t match the heights of human intel-
ligence. (After all, most of us can’t do that either.)
We should try to understand mundane thinking
first, and worry about the exceptional cases only
much later. His warning applies to AI work on cre-
ativity, too. If AI cannot simulate the rich creativi-
ty of Shakespeare and Shostakovich, it doesn’t fol-
low that it can teach us nothing about the sorts of
processes that go on in human minds—including
theirs—when people think new thoughts.
Creativity without Magic
How is creativity possible? In other words, how is
it possible for someone—or, for that matter, a com-
puter program—to produce new ideas?
At first blush, this sounds like magic: literally,
producing something out of nothing. Stage magi-
cians seem to do that, when they show us rabbits
coming out of hats. But of course it’s not really
magic at all: members of the Magic Circle know
how it’s done. In the case of creativity, the psy-
chologist—and the AI scientist—need to know
how it’s done if there’s to be any hope of modeling
it on a computer.
If we look carefully at the many examples of
human creativity that surround us, we can see that
there are three different ways in which creativity
happens. Novel ideas may be produced by combi-
nation, by exploration, or by transformation
(Boden 2004).
Combinational creativity produces unfamiliar
combinations of familiar ideas, and it works by
making associations between ideas that were pre-
viously only indirectly linked. Examples include
many cases of poetic imagery, collage in visual art,
and mimicry of cuckoo song in a classical sym-
phony. Analogy is a form of combinational cre-
ativity that exploits shared conceptual structure
and is widely used in science as well as art. (Think
of William Harvey’s description of the heart as a
pump, or of the Bohr-Rutherford solar system
model of the atom.)
It is combinational creativity that is usually
mentioned in definitions of “creativity” and that
(almost always) is studied by experimental psy-
chologists specializing in creativity. But the other
two types are important too.

Exploratory creativity rests on some culturally
accepted style of thinking, or “conceptual space.”
This may be a theory of chemical molecules, a style
of painting or music, or a particular national cui-
sine. The space is defined (and constrained) by a
set of generative rules. Usually, these rules are
largely, or even wholly, implicit. Every structure
produced by following them will fit the style con-
cerned, just as any word string generated by Eng-
lish syntax will be a gramatically acceptable Eng-
lish sentence.
(Style-defining rules should not be confused
with the associative rules that underlie combina-
tional creativity. It’s true that associative rules gen-
erate—that is, produce—combinations. But they
do this in a very different way from grammarlike
rules. It is the latter type that are normally called
“generative rules” by AI scientists.)
In exploratory creativity, the person moves
through the space, exploring it to find out what’s
there (including previously unvisited locations)—
and, in the most interesting cases, to discover both
the potential and the limits of the space in ques-
tion. These are the “most interesting” cases
because they may lead on to the third form of cre-
ativity, which can be the most surprising of all.
In transformational creativity, the space or style
itself is transformed by altering (or dropping) one
or more of its defining dimensions. As a result,
ideas can now be generated that simply could not
have been generated before the change. For
instance, if all organic molecules are basically
strings of carbon atoms, then benzene can’t be a
ring structure. In suggesting that this is indeed
what benzene is, the chemist Friedrich von Kekule
had to transform the constraint string (open curve)
into that of ring (closed curve). This stylistic trans-
formation made way for the entire space of aro-
matic chemistry, which chemists would explore
[sic] for many years.
The more stylistically fundamental the altered
constraint, the more surprising—even shocking—
the new ideas will be. It may take many years for
people to grow accustomed to the new space and
to become adept at producing or recognizing the
ideas that it makes possible. The history of science,
and of art too, offers many sad examples of people
ignored, even despised, in their lifetimes whose
ideas were later recognized as hugely valuable.
(Think of Ignaz Semmelweiss and Vincent van
Gogh, for instance. The one was reviled for saying
that puerperal fever could be prevented if doctors
washed their hands, and went mad as a result; the
other sold only one painting in his lifetime.)
Transformational creativity is the “sexiest” of
the three types, because it can give rise to ideas
that are not only new but fundamentally different
from any that went before. As such, they are often
highly counterintuitive. (It’s sometimes said that
Articles
transformation is exploration on a metalevel, so
that there’s no real distinction here [Wiggins
2006]. However, in exploratory creativity none of
the initial rules of the search space are altered,
whereas in transformational creativity some are.
We’ll see below, for example, that the style may be
varied by GAs, that is, metarules that change oth-
er rules, while remaining unchanged themselves.)
But combinational creativity is not to be sneezed
at. Kurt Schwitters and Shakespeare are renowned
for their H-creative collages and poetic images,
which depend not on stylistic transformations but
on associative processes for their origination (and
their interpretation, too). Exploratory creativity,
likewise, is worthy of respect—and even wonder.
Indeed, the vast majority of what H-creative pro-
fessional artists and scientists do involves
exploratory, not transformational, creativity. Even
Mozart and Crick and Watson spent most of their
time exploring the spaces created by their (rela-
tively rare) moments of transformation.
Despite what’s been said above, it must also be
said that there’s no clear-cut distinction between
exploratory and transformational creativity. That’s
because any rule change, however trivial, will
result in structures that weren’t possible before. So
one must decide whether to count superficial
“tweaking” as part of exploration. Since even the
average Sunday painter may make slight changes
to the style they’ve been taught, it’s probably best
to do so. And one will still have to judge, in any
given case, whether the stylistic change is superfi-
cial or fundamental.
But if creativity isn’t magic, it’s not immediately
obvious that it could be achieved or modeled by
the particular types of nonmagic offered by AI. Nor
is it immediately clear which of the three forms of
human creativity would be the easiest for AI work
to emulate, and which the most difficult.
Computer Combinations
That last question has a surprising answer. Con-
trary to what most people assume, the creativity
that’s most difficult for AI to model is the combi-
national type. Admittedly, there’s no problem get-
ting a computer to make novel combinations of
familiar (already stored) concepts. That can be
done until kingdom come. The problem, rather, is
in getting the computer to generate and prune
these combinations in such a way that most, or
even many, of them are interesting—that is, valu-
able. What’s missing, as compared with the human
mind, is the rich store of world knowledge (includ-
ing cultural knowledge) that’s often involved.
Certainly, AI programs can make fruitful new
combinations within a tightly constrained con-
text. For instance, a program designed to solve
alphabetical analogy problems, of the form If ABC
FALL 2009 25
Articles
goes to ABD, what does MRRJJJ go to? generates com-
binations that are valuable (that is, acceptable to
us as answers) and sometimes surprising, and it can
even judge one possible answer to be “better” than
another (Hofstadter 2002). In this case, for
instance, its answers included MRRJJD, MRRDDD,
MRRKKK, and MRRJJJJ; and when given XYZ
instead of MRRJJJ, it sometimes gave the surpris-
ingly elegant WYZ. Again, the chess program Deep
Blue came up with novel combinations of moves—
in one case, so seemingly uncomputerlike that
world champion Kasparov accused the computer
scientists of cheating. And programs with chemical
knowledge can come up with valuable new sug-
gestions for molecules that might have pharma-
ceutical uses.
But no current AI system has access to the rich
and subtly structured stock of concepts that any
normal adult human being has built up over a life-
time. A few systems already have access to a sig-
nificant range of concepts and factual knowledge,
stored in databases such as Wordnet, Wikipedia,
the CYC encyclopedia, and Google. And future
programs may also have increased associative and
inferential powers, based on the ontology of the
semantic web. But using huge databases sensibly,
and aptly, so as to match the combinations gener-
ated by linguistically—and culturally—sensitive
human beings is a tall order. Not impossible in
principle (after all, we don’t do it by magic), but
extremely difficult to achieve.
Consider, for instance, an example of H-creative
combinational creativity from Shakespeare. Mac-
beth, sleepless because tormented by guilt, says
this:
Sleep that knits up the ravelled sleeve of care,
The death of each day’s life, sore labour’s bath,
Balm of hurt minds, great nature’s second course,
Chief nourisher in life’s feast.
Describing sleep as someone knitting is certainly
an unfamiliar combination of two boringly famil-
iar ideas. And, assuming that Shakespeare didn’t
borrow this poetic imagery (like many of his plots)
from some previous writer, such as Petrarch, it’s H-
creative too. But why is it apt?
Well, the knitter imagined here is not producing
a new garment (the most obvious function of knit-
ting), but mending (“knitting up”) a torn sleeve.
And sleep, likewise, mends the worried mind. Sim-
ilarly, a bath (line 2) or balming ointment (line 3)
can cure the soreness caused by one’s daily work.
Moreover, in Shakespeare’s time, as in ours, the
second course of a meal (line 3) was the most nour-
ishing, the one best suited for replenishing the
body. In short, sleep provides desperately needed
relief.
It’s because we know all these things that we can
intuitively understand and appreciate Shake-
speare’s text. But a computer model needs its “intu-
26 AI MAGAZINE
ition” to be spelled out. Think what world knowl-
edge would be needed in the database even to rec-
ognize the aptness of these combinations, never
mind originating them as Shakespeare did.
Of course, AI workers could cheat. They could
build a toy system that knew only about worry-
induced sleeplessness, knitting, sleeves, ointment,
and dining customs and that—provided also with
the relevant combinational processes—could deci-
pher Shakespeare’s meaning accordingly. And even
that would be an achievement. But it wouldn’t
match the power of human minds to cope with the
huge range of creative combinations that can assail
us in a single day.
That power rests in the fact that our memories
store (direct and indirect) associations of many dif-
ferent kinds, which are naturally aroused during
everyday thinking. Shakespeare seemingly had
access to more associations than the rest of us or to
subtler criteria for judging the value of relatively
indirect associations. But our ability to understand
his poetry rests in this mundane fact about human
memory. Moreover, this fact is biological (psycho-
logical), not cultural. The specific associations are
learned within a culture, of course—as are the
socially accepted styles of thought that ground
exploratory and transformational creativity. But
the making of associations doesn’t have to be
learned: it’s a natural feature of associative memo-
ry. That’s why combinational creativity is the easi-
est of the three types for human beings to achieve.
One of the best current computer models of
combinational creativity is the joke-generating sys-
tem JAPE (Binsted, Pain, and Ritchie 1997). One
might call it a “toy” system when compared with
the human mind, but it’s more impressive than the
imaginary system just described.
JAPE’s jokes are based on combination but
involve strict rules of structure too. They are pun-
ning riddles, of a type familiar to every eight year
old. For example:
What do you call a depressed train? A low-como-
tive.
What do you call a strange market? A bizarre bazaar.
What kind of murderer has fibre? A cereal killer.
What’s the difference between leaves and a car?
One you brush and rake, the other you rush and
brake.
Hilarious, these are not. But they’re good enough
for a Christmas cracker.
Those four riddles, along with many more, were
created by JAPE. The program is provided with
some relatively simple rules for composing nine
different types of joke. Its joke schemas include:
What kind of x has y? What kind of x can y? What
do you get when you cross x with y?; and What’s
the difference between an x and a y?
The joke-generating rules are only “relatively”
simple—and much less simple than most people

would expect. That’s par for the course: AI has
repeatedly shown us unimagined subtleties in our
psychological capacities. Think for a moment of
the complexity involved in your understanding
the jest (above) about the cereal killer, and the
rather different complexities involved in getting
the point of the low-comotive or the bizarre bazaar.
Sounds and spellings, for instance, are crucial for
all three. So making (and appreciating) these rid-
dles requires you to have an associative memory
that stores a wide range of words—not just their
meanings, but also their sounds, spelling, syllabic
structure, and grammatical class.
JAPE is therefore provided with a semantic net-
work of over 30,000 units, within which new—and
apt—combinations can be made by following
some subset of the links provided. The network is
an extended version of WordNet, a resource devel-
oped by George Miller’s team at Princeton Univer-
sity and now exploited in many AI programs.
WordNet is a lexicon whose words are linked by
semantic relations such as superordinate, subordi-
nate, part, synonym, and antonym. Dimensions cod-
ing spelling, phonology, syntax, and syllable-count
had to be added to WordNet by JAPE’s programmer
so that the program could do its work, for JAPE
uses different combinations of these five aspects of
words, in distinctly structured ways, when gener-
ating each type of joke.
It wasn’t enough merely to provide the five
dimensions: in addition, rules had to be given to
enable JAPE to locate appropriate items. That is,
the rules had to define what is appropriate (valu-
able) for each joke schema. Clearly, an associative
process that obeys such constraints is very differ-
ent from merely pulling random combinations out
of the semantic network.
The prime reason that JAPE’s jokes aren’t hilari-
ous is that its associations are very limited, and
also rather boring, when compared with ours. But,
to avoid the superhuman human fallacy, we
shouldn’t forget that many human-generated jokes
aren’t very funny either. Its success is due to the
fact that its joke templates and generative schemas
are relatively simple. Many real-life jokes are much
more complex. Moreover, they often depend on
highly specific, and sometimes fleetingly topical,
cultural knowledge—such as what the prime min-
ister is reported to have said to the foreign secre-
tary yesterday. In short, we’re faced with the
“Shakespeare’s sleep” problem yet again.
Computer Exploration
Exploratory creativity, too, can be modeled by AI—
provided that the rules of the relevant thinking
style can be specified clearly enough to be put into
a computer program. Usually, that’s no easy mat-
ter. Musicologists and art historians spend lifetimes
Articles
trying to identify different styles—and they aim
merely for verbal description, not computer imple-
mentation. Anyone trying to model exploratory
creativity requires not only advanced AI skills but
also expertise in, and deep insights into, the
domain concerned.
Despite the difficulties, there has been much
greater success here than in modeling combina-
tional creativity. In many exploratory models, the
computer comes up with results that are compara-
ble to those of highly competent, sometimes even
superlative, human professionals.
Examples could be cited from, for instance,
stereochemistry (Buchanan, Sutherland, Feigen-
baum 1969), physics (Langley et al. 1987, Zytkow
1997), music (Cope 2001, 2006), architecture (Kon-
ing and Eizenberg 1981, Hersey and Freedman
1992), and visual art. In the latter category, a good
example is Harold Cohen’s program, AARON
(Cohen 1995, 2002).
AARON’s creations have not been confined to
the laboratory. On the contrary, they have been
exhibited at major art galleries around the world—
and not just for their shock value. Under develop-
ment since the late-1960s, this program has gener-
ated increasingly realistic (though not
photo-realistic) line drawings, followed by colored
images. The latter have included paintings, where-
in the paint is applied by AARON to its own draw-
ings, using paint brushes (or, more accurately,
rounded paint pads) of half a dozen different sizes.
Most recently, AARON’s colored images have been
subtly multicolored prints.
It’s especially interesting to note Cohen’s recent
remark, “I am a first-class colorist. But AARON is a
world-class colorist.” In other words, the latest ver-
sion of AARON outstrips its programmer—much as
Arthur Samuel’s checkers player, way back in the
1950s, learned how to beat Samuel himself
(Samuel 1959).
This example shows how misleading it is to say,
as people often do, “Computers can’t do anything
creative, because they can do only what the pro-
gram tells them to do.” Certainly, a computer can
do only what its program enables it to do. But if its
programmer could explicitly tell it what to do,
there’d be no bugs—and no “world-class” color
prints from AARON surpassing the handmade pro-
ductions of Cohen himself.
A scientific example—or, better, a varied group
of scientific examples—of exploratory creativity
can be found in the work of Pat Langley and
Simon’s group at CMU (Langley et al. 1987,
Zytkow 1997). This still-burgeoning set of pro-
grams is the BACON family, a dynasty, including
close relations and more distant descendants, that
has been under development since the mid-1970s
(Boden 2004, 208–222). And it is this body of
research on which Simon was relying when he
FALL 2009 27
Articles
defended AI’s status as a science (see previous text).
Among the early achievements of the BACON
family were the “discovery” of several important
laws of classical physics (Boyle’s law, Ohm’s law,
Snell’s law, Black’s law, and Kepler’s third law) and
some basic principles of chemistry (an acid plus an
alkali gives a salt; molecules are composed of iden-
tifiable elements, present in specific proportions;
and the distinction between atomic and molecular
weight). These generalizations were gleaned from
the experimental data used by human scientists
hundreds of years ago (and recorded in their note-
books); initially, the data were cleaned up for
BACON.1’s benefit, but they were later provided in
the original noisy form. The BACON suite also
reinvented the Kelvin temperature scale, by adding
a constant of 273 to the Celsius value in the equa-
28 AI MAGAZINE
© Franck Boston
tion, and “discovered” the ideal gas laws (PV / t =
k).
The words discovery and discovered need to be in
scare quotes here because this was P-creativity
rather than H-creativity. Although a few results
were historically new (for example, a version of
Black’s law that is more general than the original
one), most were not.
Later members of this set of programs were
aimed at genuine discovery, or H-creativity. Some,
for instance, suggested new experiments, intended
to provide new sets of correlations, new observa-
tions, with which the program could then work
when testing a theory. Others could introduce new
basic units of measurement, by taking one object
as the standard (human scientists often choose
water). And Simon foresaw a future in which pro-

grams modeling scientific creativity could read
papers in the scientific journals, so as to find extra
experimental data, and hypotheses, for them-
selves. (To some extent, that future has already
come: some bioinformatics software, such as for
predicting protein structure, can improve accuracy
by reading medical publications on the web. But
the ideas in more discursive scientific papers are
less amenable to AI.)
There’s an obvious objection here, however.
These programs, including the more recent ones,
all assume a general theoretical framework that
already exists. The physics-oriented BACON pro-
grams, for instance, were primed to look for math-
ematical relationships. Moreover, they were
instructed to seek the simplest relationships first.
Only if the system couldn’t find a numerical con-
stant or linear relationship (expressible by a
straight-line graph) would it look for a ratio or a
product. But one might say that the greatest cre-
ative achievement of the famous scientists mod-
eled here, and of Galileo before them, was to see
that—or even to ask whether—some observable
events can be described in terms of mathematics at
all, for this was the real breakthrough: not discov-
ering which mathematical patterns best describe
the physical world, but asking whether any math-
ematical patterns are out there to be found.
In other words, these programs were explorato-
ry rather than transformational. They were spoon-
fed with the relevant questions, even though they
found the answers for themselves. They have been
roundly criticized as a result (Hofstadter and FARG
1995, 177–179; Collins 1989), because of the
famous names (BACON and the like) used to label
them. To be sure, they can explore creatively. (And,
as remarked above, exploration is what human
chemists and physicists do for nearly all of their
time.) However, the long-dead scientists whose dis-
coveries were being emulated here did not merely
explore physics and chemistry, but transformed
them.
Could a computer ever do that?
Stylistic Transformations
Many people believe that no computer could ever
achieve transformational creativity. Given a style,
they may admit, a computer can explore it. But if
you want it to come up with a new style, don’t
hold your breath!
After all, they say, a computer does what its pro-
gram tells it to do—and no more. The rules and
instructions specified in the program determine its
possible performance (including its responses to
input from the outside world), and there’s no
going beyond them.
That thought is of course correct. But what it
ignores is that the program may include rules for
Articles
changing itself. That is, it may contain genetic algo-
rithms, or GAs (see Boden 2006, 15.vi).
GAs can make random changes in the program’s
own task-oriented rules. These changes are similar
to the point mutations and crossovers that under-
lie biological evolution. Many evolutionary pro-
grams also include a fitness function, which selects
the best members of each new generation of task
programs for use as “parents” in the next round of
random rule changing. In the absence of an auto-
mated fitness function, the selection must be made
by a human being.
Biological evolution is a hugely creative process,
in which major transformations of bodily form
have occurred. This has happened as a result of
many small changes, not of sudden saltations, and
few if any of those individual changes count as
transformations in the sense defined above. (Even
small mutations can be damaging for a living
organism, and larger—transformational—ones are
very likely to be lethal.) Nevertheless, over a vast
period of time, the evolutionary process has deliv-
ered unimaginable changes.
It’s not surprising, then, that the best prima facie
examples of transformational AI involve evolu-
tionary programming. For example, Karl Sims’s
(1991) graphics program produces varied images
(12 at a time) from which a human being—Sims, or
a visitor to his lab or exhibition space—selects one
or two for breeding the next generation. (There’s
no automatic fitness function because Sims doesn’t
know what visual or aesthetic properties to favor
over others.) This system often generates images
that differ radically from their predecessors, with
no visible family resemblance.
Sims’s program can do this because its GAs allow
not only small point mutations (leading to minor
changes in color or form) but also mutations in
which (for instance) two whole image-generating
programs are concatenated, or even nested one
inside the other. Since one of those previously
evolved programs may itself be nested, several
hierarchical levels can emerge. The result will be an
image of some considerable complexity. As an
analogy, consider these two trios of sentences:
(1) The cat sat on the mat; The cats sat on the
mat; The dog sat on the porch, and (2) The cat sat
on the mat; Aunt Flossie went into the hospital;
The cat given to me by Aunt Flossie last Christmas
before she went into the hospital in the neighbor-
ing town sat on the mat. Clearly, the second trio
displays much greater differences than the first.
So this program undeniably delivers transforma-
tions: images that are fundamentally different
from their ancestors, sometimes even from their
parents. But whether it delivers transformed styles
as well as transformed items is less clear, for family
resemblance is the essence of style. When we speak
of styles in visual art (or chemistry, or cooking), we
FALL 2009 29
Articles
mean a general pattern of ideas/artifacts that is sus-
tained—indeed, explored—over time by the artist
concerned, and perhaps by many other people too.
But Sims’s program cannot sustain a style, because
some equivalent of Aunt Flossie’s trip to the hospi-
tal can complicate the previous image at any time.
In brief, Sims’s program is almost too transfor-
mational. This lessens the importance of the selec-
tor. Even an automatic fitness function would not
prevent highly unfit examples from emerging. And
when human selectors try to steer the system
toward certain colors or shapes, they are rapidly
disappointed: sooner rather than later, unwanted
features will appear. This is frustrating for anyone
seriously interested in the aesthetics of the evolv-
ing images.
That’s why the sculptor William Latham, a pro-
fessional artist rather than a computer scientist,
uses evolutionary programming in a less radically
transformational way (Todd and Latham 1992).
His GAs allow only relatively minor changes to the
current image-generating program, such as altering
the value of a numerical parameter. Nesting and
concatenation are simply not allowed. The result
is a series of images that, he admits, he could not
possibly have imagined for himself, but that nev-
ertheless carry the stamp of his own artistic style.
The transformations, in other words, are relatively
minor and concern relatively superficial dimen-
sions of the original style (conceptual space).
It would be possible, no doubt, for an evolu-
tionary program to be allowed to make “Aunt
Flossie” mutations only very rarely. In that case,
there would be a greater chance of its producing
transformed styles as well as transformed items.
Indeed, the minor mutations might then be
regarded as exploring the existing style, whereas the
nesting/concatenating mutations might be seen as
transforming it.
Whether those stylistic transformations would
be valued is another matter. By definition, a cre-
ative transformation breaks some of the currently
accepted rules. It may therefore be rejected out of
hand—as Semmelweiss and van Gogh knew only
too well. But—as neither of them lived long
enough to find out—even if it is rejected, it may be
revived later. In biology, nonlethal mutations lead
to viable organisms, which then compete as natu-
ral selection proceeds. In human thought, social
selection takes the place of natural selection. So,
since being valuable is part of the very definition of
creative ideas, the identification of “creativity” is
not a purely scientific matter but requires socially
generated judgments.
Putatively creative ideas are evaluated by means
of a wide range of socially determined criteria. The
criteria for scientific evalutation are relatively
straightforward, and also relatively stable—even
though bitter disputes about new scientific theo-
30 AI MAGAZINE
ries can arise. Those for fashion and art are not. So
if structural transformation is necessary for a nov-
el idea to be hailed as a triumph of style-changing
creativity, it certainly isn’t sufficient.
Is Transformational
AI Actually Possible?
I said, above, that the best prima facie examples of
transformational AI involve evolutionary pro-
gramming. Why that cautious “prima facie”?
Sims’s program, after all, does generate radically
transformed images. And Latham’s program gener-
ates new visual styles, even if the family resem-
blances to the ancestor styles are relatively obvi-
ous. Moreover, we don’t need to focus only on the
contentious area of art, nor only on cases where a
human selector decides on “fitness.” Even a very
early GA program was able to evolve a sorting algo-
rithm that could put a random set of numbers into
an increasing series, or order words alphabetically
(Hillis 1992). Since then, many other highly effi-
cient algorithms have been automatically evolved
from inferior, even random, beginnings. If that’s
not transformation, what is?
Well, the objection here comes from people who
take the biological inspiration for evolutionary
programming seriously (Pattee [1985]; Cariani
[1992]; see also Boden [2006, 15.vi.c]). They
assume that AI is either pure simulation or abstract
programming that defines what sort of interac-
tions can happen between program and world (as
in computer vision, for example). And they offer a
version of the familiar argument that A computer
can do only what its program tells it to do. Specifical-
ly, they argue that genuine transformations can
arise in a system only if that system interacts pure-
ly physically with actual processes in the outside
world, as biological organisms do.
Their favorite example concerns the origin of
new organs of perception. They allow that once a
light sensor has arisen in biology, it can evolve into
better and better sensors as a result of genetic
mutations that can be approximated in AI pro-
grams. So an inefficient computer-vision system
might, thanks to GAs, evolve into a better one. But
the first light sensor, they insist, can arise only if
some mutation occurs that causes a bodily change
that happens to make the organism sensitive to
light for the very first time. The light—considered
as a physical process—was always out there in the
world, of course. But only now is it “present” for
the organism. One might say that only now has it
passed from the world into the environment.
That acceptance of light as part of the organ-
ism’s environment depends crucially on physical
processes—both in the world and in the living
body. And these processes, they say, have no place
in AI.

They grant that the generative potential of a
computer program is often unpredictable and may
even be indefinitely variable, as it is for most evo-
lutionary programs. But still, it is constrained by
the rules (including the GAs) in the program. And
if it interacts with events in the outside world, as a
computer-vision system or a process monitor does,
the types of data to which it is sensitive are preor-
dained. Certainly, they say, improved sensory arti-
facts can result from evolutionary computing. And
those improvements may be so surprising that we
naturally classify them as “transformations.” But
(so the argument goes) no fundamentally new
capacities can possibly arise.
For instance, if the physical parameters foreseen
by the programmer as potentially relevant don’t
happen to include light, then no artificial eye can
ever emerge. In general, then, there can be no real
transformations in AI.
That may be true of AI systems that are pure sim-
ulations. But it’s demonstrably not true of all AI
systems—in particular, of some work in so-called
embodied AI, for recent research in this area has
resulted in the evolution of a novel sensor: the very
thing that these critics claim can happen only in
biology.
In brief, a team at the University of Sussex were
using a GA to evolve oscillator circuits—in hard-
ware, not in simulation (Bird and Layzell 2002). To
their amazement, they ended up with a primitive
radio receiver. That is, the final (GA-selected) cir-
cuit acted as a primitive radio antenna (a “radio
wave sensor”), which picked up and modified the
background signal emanating from a nearby PC
monitor.
On investigation post hoc, it turned out that the
evolution of the radio-wave sensor had been driv-
en by unforeseen physical parameters. One of
these was the aerial-like properties of all printed
circuit boards, which the team hadn’t previously
considered. But other key parameters were not
merely unforeseen but unforeseeable, for the oscil-
latory behavior of the evolved circuit depended
largely on accidental—and seemingly irrelevant—
factors. These included spatial proximity to a PC
monitor; the order in which the analog switches
had been set; and the fact that the soldering iron
left on a nearby workbench happened to be
plugged in at the mains.
If the researchers had been aiming to evolve a
radio receiver, they would never have considered
switch order or soldering irons. Nor would either
of these matters necessarily be relevant outside the
specific (physical) situation in which this research
was done. On another occasion, perhaps, arcane
physical properties of the paint on the surround-
ing wallpaper might play a role. So we can’t be sure
that even research in embodied AI could confident-
ly aim to evolve a new sensor. The contingencies
Articles
involved may be too great, and too various. If so,
doubt about (nonaccidental) genuine transforma-
tions in AI still stands. But that they can some-
times happen unexpectedly is clear.
Computer Models
and Computer Art
All computer models of creativity are aimed at the
production of P-creative ideas, and a few at H-cre-
ativity too. And many are intended also to throw
some light on creativity in human minds. Some,
however, function in ways that have no close rela-
tion to how the the mind works: it’s enough that
they generate creative outcomes.
Examples of the latter type include most of the
AI programs employed in the various forms of
computer art. (The different types are distin-
guished, and their varying implications for “cre-
ativity” outlined, in Boden and Edmonds [2009].)
One might say that these aren’t really computer
“models” at all, but rather computer programs—
ones that may sometimes seem to work in creative
ways. (AARON was unusual in that Cohen—
already a highly successful abstract painter—first
turned to AI techniques in the hope of under-
standing his own creativity better.) Most comput-
er artists are interested not in human psychology
but in the aesthetic value of their program’s per-
formance.
That performance may be a stand-alone matter,
wherein the computer generates the result all by
itself. Having written the program, the human
artist then stands back, hands off, to let it run.
These are cases of generative art, or G-art (Boden
and Edmonds 2009).
Where G-art is involved, it’s especially likely that
the AI system itself—not just its human origina-
tor—will be credited with creativity. In evolutionary
art too (see the following text), much of the cre-
ativity may be credited to the program, for here,
the computer produces novel results—images or
melodies, for instance—that the human artist
couldn’t predict, or even imagine. In yet other cas-
es of computer art, such as the interactive art
described below, some or all of the creativity is
attributed to the programmer or the human par-
ticipants. The interactive program isn’t designed to
be (or even to appear to be) creative in its own
right, but rather to produce aesthetically attrac-
tive/interesting results in noncreative ways.
The preeminent case of G-art in the visual arts is
AARON, whose programmer tweaks no knobs
while it is running. In music, perhaps the best-
known example is the work of the composer David
Cope (2001, 2006).
Cope’s program Emmy (from EMI: Experiments
in Musical Intelligence) has composed music in
the style of composers such as Bach, Beethoven,
FALL 2009 31
Articles
Chopin, Mahler … and Scott Joplin, too. Some are
pieces for solo instrument, such as a keyboard
fugue or sonata, while others are orchestral sym-
phonies. They are remarkably compelling, striking
many musically literate listeners—though admit-
tedly not all—as far superior to mere pastiche.
That’s sometimes so even when the listener
approached Emmy’s scores in a highly sceptical
spirit. For instance, the cognitive scientist Douglas
Hofstadter, a leading figure in the computer mod-
eling of creativity (Hofstadter and FARG 1995,
Rehling 2002), believed it to be impossible that tra-
ditional AI techniques could compose music of
human quality. But on playing through some
Emmy scores for new “Chopin mazurkas,” a genre
with which Hofstadter, a fine amateur musician,
was already very familiar, he was forced to change
his mind (Hofstadter 2001, 38f.).
Other examples of computer art are not stand-
alone, but interactive (Boden and Edmonds 2009;
Boden in press); that is, the computer’s perform-
ance is continually affected by outside events
while the program is running.
Those “outside events” may be impersonal mat-
ters such as wave movements or weather condi-
tions, but usually they are the movements/actions
of human beings. Given that what the system does
on any occasion depends in part on the actions of
the human audience, the causal relation may be
obvious enough for the person to choose what
effect to have on the program’s performance.
Sometimes, however, such predictions are impos-
sible: the audience affects what happens but, per-
haps because of the complexity of the causality
involved, they don’t know how. They may not
even realize that this is happening at all—for
instance, because of a built-in delay between
(human-generated) cause and (computer-generat-
ed) effect.
One interactive program, written by Ernest
Edmonds, was chosen by the curators of a Wash-
ington, D.C., art gallery to be run alongside the
works of Mark Rothko, Clyfford Still, and Kenneth
Noland, in a 2007 exhibition celebrating the 50th
anniversary of the “ColorField” painters. (So much
for the view that computer art can’t really be art—
see below.)
An interactive artwork commissioned for the
huge millennial exhibition filling London’s new-
built Millennium Dome was described by a Times
journalist as “the best bit of the entire dome.” This
was Richard Brown’s captivating Mimetic Starfish.
The starfish is a purely virtual creature: a visual
image generated by a self-equilibrating neural net-
work that’s attached to sensors in the vicinity. The
image is projected from the ceiling down onto a
marble table, and appears to onlookers to be a large
multicolored starfish trapped inside it. But despite
being “trapped” inside the table, it moves. More-
32 AI MAGAZINE
over, it moves in extraordinarily lifelike ways, in
response to a variety of human movements and
sounds. If someone shouts, for instance, or sud-
denly pounds the table, the starfish instantly
“freezes” as a frightened animal might do.
Interactive art isn’t wholly new: Mozart’s dice
music is one ancient example. (Someone would
throw a die to decide the order in which to play
Mozart’s precomposed musical snippets, and the
result would always be a coherent piece.) But
because of the general-purpose nature of comput-
ing, a very wide range of types of interaction can
be accommodated, many of which were previous-
ly unimaginable.
In computer-based interactive art, the aesthetic
interest is not only, or not even primarily, in the
intrinsic quality of the results (images and sounds).
Rather, it is in the nature of the interaction between
computer and human beings (Boden in press). The
“audience” is seen as a participant in the creation of
the artwork—especially if the causal relations
between human activity and computer perform-
ance are direct and intelligible. In the latter case,
one can voluntarily shape the computer’s perform-
ance so as to fit one’s own preferences. But whether
the the relatively direct cases are more artisticially
interesting than the indirect ones is disputed:
there’s no consensus on just what type of interac-
tions are best from an aesthetic point of view.
As for evolutionary art, two pioneering exam-
ples (Sims and Latham) have been mentioned
already. Today, young computer artists are increas-
ingly using evolutionary techniques in their work.
One main reason is the potential for surprise that
this (randomness-based) approach provides.
Another is its connection with A-Life, and with life
itself. Some evolutionary artists even claim that
their work, or something like it, may one day gen-
erate “real” life in computers (Whitelaw 2004).
(They are mistaken, because computers—although
they use energy, and can even budget it—don’t
metabolize [Boden 1999].)
Additional types of computer art exist, which
can’t be discussed here. But there is a debilitating
occupational hazard that faces all who work in this
area, whichever subfield they focus on. Namely,
many members of the general public simply refuse
point-blank to take their work seriously.
Consider Emmy, for instance. I said, above, that
Emmy composes music in the style of Bach and oth-
er composers. I should rather have said that it com-
posed such music, for in 2005, Cope destroyed the
musical database that had taken him 25 years to
build and that stored musical features characteris-
tic of the composers concerned (Cope 2006, 364).
His reason was twofold. First, he had found over
the years that many people dismissed Emmy’s
compositions (sometimes even refusing to hear
them at all), failing to take them seriously because

of their nonhuman origin. Second, even those who
did appreciate Emmy’s scores tended to regard
them not as music but as computer output. As such,
they were seen as infinitely reproducible—and
devalued accordingly (Boden 2007). Now, howev-
er, Emmy has a finite oeuvre—as all human com-
posers, beset by mortality, do.
Some of Emmy’s detractors would be equally
adamant in dismissing every other example of
computer art. They might admit that the Mimetic
Starfish, for example, is both beautiful and—for a
while—intriguing. But they would regard it as a
decorative gimmick, not as art. For them, there can
be no such thing as computer art. Despite the fact
that there is always a human artist somewhere in
the background, the mediation of the computer in
generating what’s actually seen or heard under-
mines its status as art.
This isn’t the place to attempt a definition of the
notoriously slippery concept of art. (But remember
that a computer artwork was chosen by tradition-
al gallery curators for inclusion in Washington’s
“ColorField” exhibition; see above.) In other
words, it’s not the place for discussing whether
computer art is “really” art.
But we can’t wholly ignore a closely related
question, which is often in people’s minds when
they deny the possibility of computer art. Namely,
can a computer “really” be creative?
But Are Computers
Creative, Really?
Whether a computer could ever be “really” creative
is not a scientific question but a philosophical one.
And it’s currently unanswerable, because it
involves several highly contentious—and highly
unclear—philosophical questions.
These include the nature of meaning, or inten-
tionality; whether a scientific theory of psycholo-
gy, or consciousness, is in principle possible; and
whether a computer could ever be accepted as part
of the human moral community. Indeed, you can
ignore creativity here, for many philosophers argue
that no naturalistic explanation of any of our psy-
chological capacities is possible, not even an expla-
nation based in neuroscience. In short, the philo-
sophical respectability of “strong” AI, and of
cognitive science in general, is hotly disputed.
These are among the very deepest questions in
philosophy. I’ve discussed them elsewhere (Boden
2004, chapter 11; Boden 2006, chapter 16). I’ve
also argued that the ultrasceptical, postmodernist
view is irrational and fundamentally self-defeating
(Boden 2006, 1.iii.b, 16.vi–viii). But there’s no
knock-down refutation on either side. That being
so, even the youngest readers of AI Magazine
shouldn’t expect to see these questions to be defin-
itively answered in their lifetimes.
Articles
The scientific questions offer more hope.
Enough advance has already been made in com-
putational psychology and computer modeling to
make it reasonable to expect a scientific explana-
tion of creativity. Optimists might even say that it’s
already on the horizon. This doesn’t mean that
we’ll ever be able to predict specific creative ideas,
any more than physicists can predict the move-
ments of a single grain of sand on a windswept
beach. Because of the idiosyncracy and (largely
hidden) richness of individual human minds, we
can’t even hope to explain all creative ideas post
hoc. But, thanks in part to AI, we have already
begun to understand what sort of phenomenon cre-
ativity is.
Still something of a mystery, perhaps. And cer-
tainly a marvel. But not—repeat, not—a miracle.
References
Binsted, K.; Pain, H.; and Ritchie, G. D. 1997. Children’s
Evaluation of Computer-Generated Punning Riddles.
Pragmatics and Cognition 5(2): 305–354.
Bird, J., and Layzell, P. 2002. The Evolved Radio and Its
Implications for Modelling the Evolution of Novel Sen-
sors. In Proceedings of Congress on Evolutionary Computa-
tion, CEC-2002, 1836–1841. Pisacataway, NJ: Institute of
Electrical and Electronics Engineers.
Boden, M. A. In press. Aesthetics and Interactive Art. In
Art, Body, Embodiment, ed. R. L. Chrisley, C. Makris, R. W.
Clowes, and M. A. Boden. Cambridge: Cambridge Schol-
ars Press.
Boden, M. A. 2007. Authenticity and Computer Art. Spe-
cial Issue on Creativity, Computation, and Cognition, ed.
P. Brown. Digital Creativity. 18(1): 3–10.
Boden, M. A. 2006. Mind as Machine: A History of Cogni-
tive Science. Oxford: Oxford University Press.
Boden, M. A. 2004. The Creative Mind: Myths and Mecha-
nisms, 2nd ed. London: Routledge.
Boden, M. A. 1999, Is Metabolism Necessary? British Jour-
nal for the Philosophy of Science 50(2): 231–248.
Boden, M. A., and Edmonds, E. A. 2009. What Is Gener-
ative Art? Digital Creativity 20(1–2): 21–46.
Buchanan, B. G.; Sutherland, G. L.; and Feigenbaum, E.
A. 1969, Heuristic DENDRAL: A Program for Generating
Explanatory Hypotheses in Organic Chemistry. In
Machine Intelligence 4, ed. B. Meltzer and D. M. Michie,
209–254. Edinburgh: Edinburgh University Press.
Cariani, P. 1992. Emergence and Artificial Life. In Artificial
Life II, ed. C. J. Langton, C. G. Taylor, J. D. Farmer, and S.
Rasmussen, 775–797. Reading, MA: Addison-Wesley.
Cohen, H. 1995. The Further Exploits of AARON Painter.
In Constructions of the Mind: Artificial Intelligence and
the Humanities, ed. S. Franchi and G. Guzeldere. Special
edition of Stanford Humanities Review, 4(2): 141–160.
Cohen, H. 2002. A Million Millennial Medicis. In Explo-
rations in Art and Technology, ed. L. Candy and E. A.
Edmonds, 91–104. London: Springer.
Collins, H. M. 1989. Computers and the Sociology of Sci-
entific Knowledge, Social Studies of Science, 19: 613–624.
Cope, D. 2006. Computer Models of Musical Creativity.
Cambridge, Mass.: The MIT Press.
FALL 2009 33
Articles
ACM IUI'10
Conference on Intelligent User Interfaces
Hong Kong, China; 7-10 February 2010
IUI is where the Human-Computer Inter-
action (HCI) community and the Artificial
Intelligence (AI) community meet.
Conference Co-Chairs:
Charles Rich (Worcester Polytechnic Inst., USA);
Qiang Yang (Hong Kong U. of Sci. & Tech.,
China)
Program Co-Chairs:
Marc Cavazza (U. Teesside, UK);
Michelle Zhou (IBM Research, China)
For more information, the conference website is:
http://www.iuiconf.org/
(in cooperation with AAAI)
Cope, D. 2001. Virtual Music: Computer Synthesis of Musi-
cal Style. Cambridge, Mass.: The MIT Press.
Hersey, G., and Freedman, R. 1992. Possible Palladian Vil-
las Plus a Few Instructively Impossible Ones. Cambridge,
MA: The MIT Press.
Hillis, W. D. 1992. Co-Evolving Parasites Improve Simu-
lated Evolution as an Optimization Procedure. In Artifi-
cial Life II, ed. C. J. Langton, C. G. Taylor, J. D. Farmer,
and S. Rasmussen, 313–324. Reading, MA: Addison-Wes-
ley.
Hofstadter, D. R. 2002. How Could a COPYCAT Ever be
Creative? In Creativity, Cognition, and Knowledge: An Inter-
action, ed. T. Dartnall, 405–424. London: Praeger.
Hofstadter, D. R., and FARG, The Fluid Analogies
Research Group. 1995. Fluid Concepts and Creative Analo-
gies: Computer Models of the Fundamental Mechanisms of
Thought. New York: Basic Books.
Hofstadter, D. R. 2001. Staring Emmy Straight in the
Eye—And Doing My Best Not to Flinch. In Virtual Music:
Computer Synthesis of Musical Style, ed. D. Cope, 33–82.
Cambridge, Mass.: The MIT Press.
Koning, H., and Eizenberg, J. 1981. The Language of the
Prairie: Frank Lloyd Wright’s Prairie Houses. Environment
and Planning, B, 8(4): 295–323.
Langley, P. W.; Simon, H. A.; Bradshaw, G. L.; and Zytkow,
34 AI MAGAZINE
View publication stats
J. M. 1987. Scientific Discovery: Computational Explorations
of the Creative Process. Cambridge, MA: The MIT Press.
Todd, S. C., and Latham, W. 1992. Evolutionary Art and
Computers. London: Academic Press.
Lenat, D. B. 1983. The Role of Heuristics in Learning by
Discovery: Three Case Studies. In Machine Learning: An
Artificial Intelligence Approach, ed. R. S. Michalski, J. G.
Carbonell, and T. M. Mitchell, 243–306. Palo Alto, CA:
Tioga.
McCarthy, J.; Minsky, M. L.; Rochester, N.; and Shannon,
C. E. 1955. A Proposal for the Dartmouth Summer
Research Project on Artificial Intelligence. AI Magazine
27(4): 12–14.
Pattee, H. H. 1985. Universal Principles of Measurement
and Language Functions in Evolving Systems. In Com-
plexity, Language, and Life: Mathematical Approaches, ed. J.
Casti and A. Karlqvist, 168–281. Berlin: Springer-Verlag.
Rehling, J. A. 2002. Results in the Letter Spirit Project. In
Creativity, Cognition, and Knowledge: An Interaction, ed. T.
Dartnall. London: Praeger, 273–282.
Samuel, A. L. 1959. Some Studies in Machine Learning
Using the Game of Checkers. IBM Journal of Research and
Development 3: 211–229.
Simon, H. A. 1995. Explaining the Ineffable: AI on the
Topics of Intuition, Insight, and Inspiration. In Proceed-
ings of the Fourteenth International Joint Conference on Arti-
ficial Intelligence, vol. 1, 939–948. San Francisco: Morgan
Kaufmann Publishers.
Sims, K. 1991. Artificial Evolution for Computer Graph-
ics. Computer Graphics. 25(4): 319–28.
Whitelaw, M. 2004. Metacreation: Art and Artificial Life.
London: The MIT Press.
Wiggins, G. 2006. Searching for Computational Creativ-
ity. New Generation Computing 24(3): 209–222.
Zytkow, J. ed. 1997. Machine Discovery. London: Kluwer
Academic.
Margaret A. Boden (research professor of cognitive sci-
ence at the University of Sussex) was the founding dean
of Sussex University’s School of Cognitive and Comput-
ing Sciences. She is a Fellow (and former vice-president)
of the British Academy, a member of the Academia
Europaea, a Fellow of the Association for the Advance-
ment of Artificial Intelligence, a Fellow of the European
Coordinating Committee for Artificial Intelligence, a Life
Fellow of the UK’s Society for Artificial Intelligence and
the Simulation of Behaviour, a member of Council of the
Royal Institute of Philosophy, and the former chairman
of Council of the Royal Institution of Great Britain.
Boden has lectured widely, to both specialist and general
audiences, in North and South America, Europe, India,
the former USSR, and the Pacific. She has also appeared
on many radio and TV programs in the UK and else-
where. Her work has been translated into 20 foreign lan-
guages. She was awarded an OBE in 2001 (for “services to
cognitive science”), and beside her Cambridge ScD, she
has three honorary doctorates (from Bristol, Sussex, and
the Open Universities).